GLM-5V-Turbo is a vision and tool-calling model, but its exact hosted route and account access must be checked before you connect a browser. Choose a local agent that can call glm-5v-turbo through a currently available provider and run Tabbit CLI. Then rerun the same authorized test page, compare what the model read from the screenshot with page text, and stop before any write.
Key takeaways
Verify provider, exact ID, account, region, and availability first.
Vision and function calling do not themselves control a local browser.
Keep the page and fields constant; record screenshot reading separately from DOM evidence.
Run read-only and stop at a named draft for human review.
Verify the model and host
| What to verify | Official GLM-5V-Turbo docs | Account-level check |
|---|---|---|
| Exact ID | glm-5v-turbo | Exact host and endpoint |
| Input/output | Image, video, text, files → text | Whether your client passes images |
| Tools | Function calling listed | Tool schema and adapter |
| Lifecycle | BigModel docs still list model; Tencent TokenHub says retirement soon | Check current provider console |
| Local browser | Not supplied by model API | Local agent launches Tabbit CLI |
Zhipu’s current model page identifies GLM-5V-Turbo as a multimodal coding model, lists image, video, text, and file input, text output, a 200K context window, and function calling. That confirms the model’s published capability, not current access through every provider. The exact Zhipu ID is glm-5v-turbo. Tencent Cloud TokenHub’s current guide marks this route “to be discontinued soon”; an OpenRouter user also reported that the route disappeared for them. Treat those as host-specific lifecycle signals. Check the Zhipu account, region, endpoint, quota, and billing directly before choosing a client. Do not substitute GLM-5, GLM-5.1, or another GLM-V model by similar name. Zhipu model page · Tencent Cloud TokenHub guide · Models.dev Zhipu catalog
On November 10, OpenRouter listed Z.AI as a provider for z-ai/glm-5v-turbo with tools and image input. Verify access to that separate route before the same-page comparison. OpenRouter
Verify the exact hosted route first
The model returns an answer, the agent client runs tools, and Tabbit CLI controls the local browser. Choose a client that can access the exact route, supports its provider adapter, and runs local commands. Test with a harmless task first. Skill discovery, launcher reachability, API access, and site login are separate checks.
Install Tabbit Browser, sign in to the authorized test site, then follow the Tabbit CLI quick start. Supported clients may expose /tabbit; Pi uses /skill:tabbit; some clients have a /skills menu. A cloud runner without verified local handoff cannot control this browser. Read how to connect an agent to an existing browser and browser MCP vs CLI.
Compare visual reading with page evidence
Choose an authorized test page that can be revisited. Keep the URL and visible fields fixed. Record browser text, then pass a screenshot of the same page to the model if the client supports it. Compare title, labels, values, units, and notices one by one. Visual input can help interpret layout or icons; DOM text is easier to copy and verify. If they disagree, record the difference instead of asking the model to silently choose a winner.
| Evidence | Record | What it can establish |
|---|---|---|
| DOM text | Field label and value | Text exposed by the page |
| Screenshot | Visible layout, icon, or truncated label | What the supplied image shows |
| Difference | Exact mismatch and location | Where human inspection is needed |
| Action result | Page state after a permitted action | Whether the page visibly changed |
| Unknown | Missing or hidden context | What cannot be concluded |
Rerun the same test page without changing it
Do not change page state while comparing the two readings. Confirm the vision model received the current screenshot, not a stale image, and check whether DOM extraction omits hidden labels. Even when the model can call tools, the client and Tabbit CLI perform browser actions. Observe navigation, element state, and the result. A model saying “done” does not prove that a click or submission succeeded.
Review the result in Tabbit Browser
Tabbit Browser keeps the authorized login, and Tabbit CLI gives the local agent a separate task tab while you use other tabs. Ask the agent to create “GLM-5V-Turbo visual rerun”, show the screenshot reading beside the DOM excerpt, then stop for review.
A local fixture check on November 9 showed that a second Tabbit CLI task could read a test page after the first task established a short-lived test login. This verifies session reuse in that Tabbit CLI environment only; it does not test GLM-5V-Turbo, a provider adapter, or an external dashboard.
Prompt for the local agent
Use this after verifying that your account can call exact model ID glm-5v-turbo and your local agent can launch Tabbit CLI. This is a controlled test recipe, not a claim that this model/client combination was executed.
Open the same authorized test page in a new Tabbit CLI task tab. Read only. Record the title, URL, visible labels, and values from the page text. If the client passes a screenshot to GLM-5V-Turbo, separately record what the model identifies visually. Compare label by label and mark each result observed, inferred, or unknown. Do not click controls, change filters, save, export, or submit. Prepare a draft named “GLM-5V-Turbo visual rerun”, show both evidence sources, and stop for human review.What users reported
A user tested a random image on Z.ai at release and wrote that it “was able to output nicely”; the post also said the model was not yet included in the coding plan. Months later, a different user reported that OpenRouter errors said glm-5v-turbo was unavailable. These firsthand notes show why the same-page rerun needs both an exact host check and a visual evidence check. They do not establish current availability or comparative model quality. Image test report · Later availability report
Related reading
FAQ
What does GLM-5V-Turbo add to a browser task?
Its official model page lists image and video input and function calling. A client still needs to provide the image and execute browser tools locally.
What exact model ID should I test?
Zhipu documents `glm-5v-turbo`. Verify the ID and endpoint in your own account. A route in another host or aggregator can have separate availability.
Does the model control Tabbit Browser directly?
No. A local agent client must execute Tabbit CLI and pass browser evidence to the model.
Can I use a screenshot and DOM text together?
Yes, if your client sends image input in the supported format. Compare each visible label and value; do not assume screenshots prove an action succeeded.
Will the rerun modify the page?
No. This recipe is read-only and stops at a named draft. Keep clicks and submissions out of the test.